Control of Discrete-Time PartiallyObserved Jump Linear Systems Over Causal Communication Systems C. D. Charalambous Depart. of ECE University of Cyprus Nicosia, Cyprus CDC 2006, San Diego S. Z. Denic Depart. of ECE University of Arizona Tucson 1 Control Over Communication Channel Block diagram of a control-communication problem The source is partially observed jump system and communication channel is causal Dynamical System Sensor Encoder Collection and Transmission of Information (Node 1) CDC 2006, San Diego Communication Channel Capacity Limited Link Decoder Sink Reconstruction with Distortion Error (Node 2) 2 Critical Features Critical features from the communication and control point of view Amount of data produced by a particular source (sensor) – source entropy Capacity of the communication channel Controllability and observability of the controlled system CDC 2006, San Diego 3 References [Plenty More] Nair, Dey, and Evans,“Communication limited stabilisability of jump Markov linear systems,” In Proc. 15th Ini. Symp.Math. The. New. Sys., U. Notre Dame, USA, Aug 2002. Nair, Dey, and Evans, “Infimum data rates for stabilising Markov jump linear systems,” in Proc. 42th IEEE Conf Dec. Contr., pp. 1176-1181, 2003. C. D. Charalambous, “Information theory for control systems: causality and feedback,” in Workshop on Communication Networks and Complexity, Athens, Greece, August 30-September 1, 2006. CDC 2006, San Diego 4 Overview Problem formulation Causal communication channels and systems Mutual information for causal channels Data processing inequalities for causal communication channels Capacity for causal communication channels Rate distortion for causal communication channels Information transmission theorem Necessary conditions for observability and stabilizability over causal communication channels Conclusions CDC 2006, San Diego 5 Problem Formulation Problem formulation Causal communication channels and systems Necessary conditions for observability and stabilizability over causal communication channels Conclusions CDC 2006, San Diego 6 Problem Formulation Block diagram of control/communication system X t 1 A St X t B St Wt N St U t , X 0 X Yt C St X t D St Vt St t 0 , St 1,..., M Pr St 1 j | St i pij Wt N 0, I k Vt N 0, I l X 0 N x0 , Q0 Wt ,Vt , St , X 0 : t N 0 CDC 2006, San Diego independent 7 Problem Formulation Encoder, Decoder, Controller are causal Z t c Y0t , Z 0t 1, Z 0t 1, S0t Yt d Z 0t , Y0t 1, S0t U t Y0t ,U 0t 1, S0t with feedback Communication channel causality P dZ t | z0n , z0t 1 P dZ t | z0t , z0t 1 , n t CDC 2006, San Diego 8 Problem Formulation System performance measures Definition 2.1: (Observabilit in Probability). The system is observable in probability if for any D, δ ≥ 0 there exist an encoder and decoder such that 1 t 1 lim Pr Yk Yk D, t t k 0 Definition 2.2: (Observability in r-th mean). The system is observable in r-th mean if there exist an encoder and decoder such that r 1 t 1 lim E Yk Yk D, r 0 t t k 0 where D ≥ 0 is finite. CDC 2006, San Diego 9 Problem Formulation System performance measures Definition 2.3: (Stabilizability in probability). The system is stabilizable in probability if for any D, δ ≥ 0 there exist a controller, encoder and decoder such that 1 t 1 lim Pr X k 0 D, t t k 0 Definition 2.4: (Stabilizability in r-th mean). The system is asymptotically stabilizable in r-th mean if there exist a controller, encoder and decoder such that 1 t 1 r lim E X k 0 D, r 0 t t k 0 where D ≥ 0 is finite. CDC 2006, San Diego 10 Causal Communication Channels and Systems Problem formulation Causal communication channels and systems Mutual information for causal channels Data processing inequalities for causal communication channels Capacity for causal communication channels Rate distortion for causal communication channels Information transmission theorem Necessary conditions for observability and stabilizability over causal communication channels Conclusions CDC 2006, San Diego 11 Causal Communication Channels and Systems Lemma 3.2: Let iR Z 0T 1; Z 0T 1 log p z0T 1 | z0T 1 p z0T 1 denote the self-mutual information when the RND p z0T 1 | z0T 1 p z0T 1 is restricted to a non-anticipative or causal feedback channel with memory. Then, the restricted mutual information is given by T T 1 T 1 T 1 T 1 I Z i ; Z | Z i 1 IC Z0 ; Z0 E i R Z 0 ; Z 0 0 i 0 i 0 Directed Information CDC 2006, San Diego 12 Causal Communication Channels and Systems Remark: In general, causal mutual information is not symmetric IC Z0T 1; Z0T 1 I C Z 0T 1; Z 0T 1 Data processing inequality for causal channels I Z 0n ; Z 0n I C Z 0n ; Z 0n I Y0t ; Z 0n I Y0t ;Y0t I C Y0t ;Y0t , n, t N 0 CDC 2006, San Diego 13 Causal Communication Channels and Systems Channel capacity based on the causal mutual information 1 1 CC lim CT lim sup IC Z0T ; Z 0T T T T T p T Z 0 Rate distortion based on the causal mutual information 1 1 RT D lim inf I C Y0T ;Y0T T T T T p T T M Y0 |Y0 RC D lim CDC 2006, San Diego 14 Causal Communication Channels and Systems Theorem 4.1: (Information Transmission Theorem) Suppose the different communication blocks in Fig. 1 form a Markov chain. Consider a control-communication system where the communication channel is restricted to being causal. A necessary condition for reconstructing a source signal Yt up to a distortion level D from Z t is given by RC D CC CDC 2006, San Diego 15 Necessary conditions for observability and stabilizability over causal communication channels Problem formulation Causal communication channels and systems Mutual information for causal channels Data processing inequalities for causal communication channels Capacity for causal communication channels Rate distortion for causal communication channels Information transmission theorem Necessary conditions for observability and stabilizability over causal communication channels Conclusions CDC 2006, San Diego 16 Necessary conditions for observability and stabilizability over causal communication channels Lemma 4.2. Consider the following single letter distortion measure 1 T (Yi Yi ) : R p [0, ) T (Y0T ,Y0T ) (Yi Yi ) , where T i 0 Then, a lower bound for 1 RT ( D ) T is given by 1 1 RT ( D) H S (Y0T ) max H S (h), T T hGD p where GD {h : R [0, ); h( y )dy 1, ( y )h( y )dy D}. Rp Rp It follows RC ( D ) ( 0 ) H S (h* ) and under some conditions, this lower bound is exact for D 0. CDC 2006, San Diego 17 Necessary conditions for observability and stabilizability over causal communication channels Theorem 4.3. Consider a jump control-communication system where Yt R p is the observed process at time t. Let pS s be a steady state distribution of the underlying Markov chain. Introduce the following notation X t E X t | Y0t 1,U0t 1, S0t 1 Q lim Qt Xt Qt E X t X t tr | Y0t 1,U 0t 1, S0t 1 t A necessary condition for asymptotic observability and stabilizability in probability is given by CC Mp 1 log 2 e log[(2 e) p det g ] 2 2 log det C i Q,i C tr i D i D tr i p i S i CDC 2006, San Diego 18 Necessary conditions for observability and stabilizability over causal communication channels g is the covariance matrix of the Gaussian distribution h* ( y ) ~ N (0, g ),( y R p ) which satisfies || y|| h* ( y )dy D. A necessary condition for asymptotic observability and stabilizability in r-th mean is given by p Mp CC log 2 e log e r log( 2 p p ( ) r ). r p dVd ( ) rD r log det C i Q,i C tr i D i D tr i p i S i CDC 2006, San Diego 19 Conclusion General necessary conditions for observability and stabilizability for jump linear systems controlled over a causal communication channel are derived. Causal Information Theory is Essential for Channels with Feedback and Memory Different criteria for observability and stabilizability corresponds to different necessary condition. Future work Sufficient conditions (design encoders and decoders) Channel-source matching CDC 2006, San Diego 20